Choporova O. Damage zones forecasting for engineering constructions using machine learning

The methods and models of computer training developed in the dissertation allow to make quick estimates of the parameters of the state of the object, namely the determination of the stress-strain state. During the dissertation research, artificial neural networks were developed to determine the maximum deflection and intensity of Mises stresses in plates and shells. Algorithms for sampling for training and testing of models were also developed. Possibilities of application of genetic algorithm for optimization of a neural network of regression analysis and forecasting of the maximum deflection of plates and covers are investigated. The following results were obtained: − for the first time a neural network method was developed to determine the stress-strain state of plates with arbitrary boundary conditions, which allows to determine the maximum deflection, and the maximum value of the stress intensity according to Mises; − for the first time a neural network method was developed to determine the stress-strain state of a cylindrical shell and a combination of cylindrical and conical shells, which allowed to determine the deflection and intensity of stresses according to Mises; − for the first time developed a neural network method based on the architecture "autocoder" to determine possible fracture zones in square plates with a cut-out by generating probable patterns of stress distribution; − improved neural network methods for modeling the stress-strain state of structures by developing a genetic algorithm for optimizing the architecture of the neural network, which allowed to increase the accuracy of prediction. The software implementation of the developed methods is performed in the Python programming language using the scikit-learn, numpy, Pandas and Keras libraries.